Está en la página 1de 9

International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and


Geoinformation
journal homepage: www.elsevier.com/locate/jag

Application of Landsat 8 imagery to regional-scale assessment of lake


water quality
Jacek Andrzej Urbanski a,∗ , Agnieszka Wochna a , Iwona Bubak b , Waldemar Grzybowski b ,
Katarzyna Lukawska-Matuszewska b , Magda Łacka ˛ d
, Sylwia Śliwińska b ,
Bożena Wojtasiewicz b,c , Marek Zajaczkowski
˛ d

a
GIS Centre, University of Gdansk, 80-952 Gdansk, Poland
b
Institute of Oceanography, University of Gdansk, 81-378 Gdynia, Poland
c
CSIRO Oceans & Atmosphere, Floreat, WA, Australia
d
Institute of Oceanology Polish Academy of Sciences, 81-712 Sopot, Poland

a r t i c l e i n f o a b s t r a c t

Article history: The aim of the project was to create a tool with which to support regional lake quality assessment using
Received 29 November 2015 Landsat 8 imagery data. The model of assigning the ecological status was implemented in GIS for the
Received in revised form 6 April 2016 northern part of Poland and classifies lake quality for several classes according to classification of WFD
Accepted 18 April 2016
using two basic assumptions. The first is that there exists a combination of OLI bands (OLI2/OLI4 was used)
which correlates well with the trophic state of the lakes; the second assumption is that the reference
Keywords:
trophic state depends on the mean depth of the lake. The model uses a lake geodatabase which contains
Landsat 8
lakes outlines, raster masks of lakes and attribute information about their mean depth. There is no need
Lakes
Trophic status
to provide any field data when using this tool, as calibration of the model is done using subsets of lakes
GIS which were classified using legally defined methods. The tool allows fast classification of 2800 lakes from
Water quality the area of interest. The results show good agreement between satellite and expert based methods.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction Such places may be found in Canada, Northern US, Sweden, Finland,
Poland and Baltic Countries. They have a common origin connected
Lake water quality is of a global interest. Indeed, the high or with the end of the last glaciation. They are surrounded by a mix-
bad ecological state of rivers, lakes and coastal waters can strongly ture of land cover and use. Some lake watersheds constitute a nearly
influence the values of properties along their coasts (McCullough unchanged environment, while others are strongly affected by agri-
et al., 2012). In nearly every country it is the responsibility of local cultural runoff, erosion, urban development and loss of wetland,
administration to care for the health of the water bodies. In the EU, all of which have a strong impact on lake health (Torbick et al.,
the WFD (Water Framework Directive) obligates all member coun- 2013). This is the main reason why the trophic status of lakes incor-
tries to achieve good ecological quality of all inland water bodies porating relationships between nutrients (mainly phosphorus),
until 2015 (Peeters et al., 2009). These demands require the estab- phytoplankton and transparency may vary from ultra-oligotrophic
lishment of complex monitoring and assessment systems. They are to hyper-eutrophic (Solimini et al., 2006). On the other hand how-
traditionally based on water sampling at fixed stations and labora- ever, the trophic status of lakes with minimal human influence may
tory analyses. This method may be difficult to implement when also vary in a natural way depending on the lake morphometry, and
there are hundreds of separate water bodies to monitor (Kalio, especially with regard the mean depth of the water body (Scheffer,
2000). There are several spots in the northern hemisphere where 1998).
the lakes are clustered and their number exceeds the thousands. There are many parameters used in lake monitoring which typi-
cally include measures of chlorophyll-a (Chl-a), suspended material
(SPM), light attenuation for which the common proxy is Secchi
depth (SD) and colored dissolved organic matter (CDOM). The WFD
∗ Corresponding author.
requirements are extremely demanding and aim to assess five qual-
E-mail address: cgisju@ug.edu.pl (J. Andrzej Urbanski).

http://dx.doi.org/10.1016/j.jag.2016.04.004
0303-2434/© 2016 Elsevier B.V. All rights reserved.
J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36 29

ity classes (Bad, Poor, Moderate, Good, High) by comparing the depth related classes (Phillips et al., 2008). The tool should make it
current status to reference state using biological, physical as well possible to calibrate the classification process in an iterative way
chemical indicators supplemented by expert knowledge. However, so as the subset of lakes whose quality state is known may be used
some publications suggest that this may be predicted quite well to calibrate the complete set of lakes.
using water transparency only expressed by Secchi depth (Peeters
et al., 2009).
The remote sensing monitoring of lakes primarily uses satellite 2. Data and methods
data, mainly from Landsat TM and TM+ sensors which have the per-
fect spatial and time resolution for such an aim. Indeed, there are The development of the system was carried out in two steps.
many examples of when this tool has been used successfully, and First the two field experiments were conducted using two Landsat
these can be divided into two groups of applications. The first group 8 scenes and field observations carried out at nearly the same time
uses a singular or multiple images of the area of interest and field as the images were taken. The aim of these experiments was to find
data obtained from several lakes (Torbick et al., 2013; Tebbs et al., out the best combination of satellite image bands for describing the
2013; Duan et al., 2008; Sass et al., 2007; Brezonik et al., 2005). It has trophic state of the lakes in the area of interest and to prove that
been suggested in several publications that the time gap between our assumption about the role of mean depth is true. On the basis
lake data collection and satellite data should be no longer then of the results from these two experiments the model of assigning
3–10 days (Olmanson et al., 2008; McCullough et al., 2012). In most the ecological status using only satellite data was developed and
cases the band ratio multi regression approach is used to obtain the verified using independent expert assessments.
relation between satellite signal and in-situ measurements. Both
raw signal (DN), radiance and reflectance have been used. Sev- 2.1. Description of study area
eral methods of less or more advanced atmospheric corrections
were applied and discussed, although there are suggestions that The study area covers the northern part of Poland, where there
simple dark object subtraction is recommended for classification are many densely distributed clusters of postglacial lakes. There are
and change detection applications (Song et al., 2001). The results approximately 8000 lakes with a total surface of around 3000 km2
of these projects are, general speaking, similar. They have proved across, and a total area of approximately 116000 km2 (Fig. 1B). The
that it is possible to obtain good results for the mapping of SD mean depth of the lakes varies from 0.2–38.7 m. The land cover of
and chlorophyll-a, as well as weaker results for SPM and CDOM, this area is a mosaic of agricultural areas (61%), forests and semi
cyanobacteria or diatoms. In order to obtain acceptable results, the natural areas (34%) as well as urban and suburban fabric (2%). The
regression equations have to be calibrated separately for a partic- field experiment area (Fig. 1) is placed in the central area of study
ular image. The second group used monitoring programs and the which is representative of climate and land cover. In Poland, the
community approach to collect Secchi depth data from many lakes ecological state of the lakes (required by WFD) is determined by
over several years (McCullough et al., 2012; Olmanson et al., 2008; the regional inspectorates of environmental protection which are
Kloiber et al., 2002). They also used the regression approach and the branches of the government administration. They use several
conducted additional analyses including change detection. legally defined procedures in order to divide the ecological status
The use of satellite images for lake monitoring, especially in of lakes into five classes using mainly biological based methods
places with hundreds or thousands of lakes, is very tempting. supplemented by some abiotic measurements. These methods are
Indeed, recent years have seen the emergence of two important expensive and need a lot of laboratory work. This means that a
facts. The first is there are now new satellites platforms whose limited number of lakes may be classified each year. For example,
resolution is suitable for lake monitoring, and the second is that only 22 lakes were assessed in 2014 in the Pomeranian District
their usage is free of charge. Using both Landsat 8 and, in the near (Fig. 1).
future, Sentinel 2, it will be possible to obtain at least one uncloudy
image nearly every month. This will give environmental managers
a chance to use all benefits of RS technology such as simultaneously 2.2. Lake geodatabase
monitoring many lakes.
The aim of the project was to create a GIS tool to support regional The geodatabase of lakes for the northern part of Poland was cre-
lake quality assessment. The tool should classify lake quality for ated using three sources: A vector layer from polish hydrographic
classes according to classification used by WFD. We assumed that geodatabase MPHP in the form of polygons (SHP) containing lake
the tool will require only Landsat 8 image which covers a few dozen names; an analogous atlas of lakes (Janczak, 1996) which contains
of lakes, the water mask of the lakes and its mean depth. The tool tables with positions of the central points of the lakes, and their
will not require any field data which will make them more flexible mean and maximum depth; and Landsat 8 images. First the max-
and operational. To fulfill these aims three research problems have imum depth from atlas tables was spatially joined with shapes of
to be solved. First, the best method has to be found out for estima- the MPHP lakes. Following this, the number of lakes was updated
tion of crucial parameters used in remote sensing lake monitoring. as the MPHP geodatabase contains many small lakes which are not
Then, such a model has to be created which allows to compare currently in existence and which are now covered by vegetation.
the current ecological status of the lakes using the results of satel- This was done using Landsat 8 images for the whole area of northern
lite image analyses. Finally it has to be found out how to calibrate Poland. The area (Fig. 1B) was covered by twelve uncloudy images
the results of this comparison using the lakes with known current (paths: 187–192 and rows: 22–24) from the years 2013–2015. The
state or defined as reference state lakes. The mean depth of the shortwave infrared band (6th band DN) of Landsat 8, which gives
lake plays an important role. The postglacial lakes may have dif- the strongest and most stable difference between water and land,
ferent depths on which their ecological state in natural conditions and the median statistics were used to find a threshold to select
is heavily dependent. There is a well-established difference in the existing lakes only. The same set of images was used to create a
structure and functioning of the lakes which depends on their mean raster of lake zones for extracting pixel values for statistics calcula-
depth (Scheffer, 1998; Poikane et al., 2014). The natural ecological tion. It was decided that for each lake only pixels with values lower
state is different in polymictic shallow systems than in deeper sea- than the median of all lake pixels in band 6 would create a zone
sonally stratified systems. As a result, and in order to minimize with unique lake identifiers. As a result, our geodatabase contains
natural biological variations, lakes are often divided into arbitrary 2800 lakes with the same mean and maximum depth as a vector
30 J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36

Fig. 1. Study area. A—Landsat 8–field experiment area with trophic state classes of sampled lakes determined during experiment and sample stations; B—area covered by
Lake Geodatabase and model of assessments of lake water quality (red rectangle shows position of field experiment area, while orange outline is a border of Pomerania
District).

polygon layer and a lake sampling zone raster layer; both have a through analysis, as described in Section 2.5 (please see Fig. 1A). The
unique lake identifier (LID). field measurements were carried out on 6–8 June and 11–13 July,
meaning that all samples were collected no later than 6 days after
2.3. Field experiment the satellite images were taken. The field measurements included:
DGPS position, Secchi depth (SD), chlorophyll-a (Chl-a), suspended
The satellite part of the field experiments was carried out on 5 particulate matter (SPM), colored dissolved organic matter (CDOM)
June (Scene 1) and 7 July (Scene 2) 2014. The images were partly and dissolved organic carbon (DOC). The composition of the phyto-
cloudy but free from haze and cirrus contamination. First, on the plankton species and its biomass were also determined in order
same day as the Landsat 8 images were taken, the scenes were con- to assess the seasonal difference in the development of various
verted to natural color composite and visually inspected to select planktonic algal groups between both scenes. Secchi depth was
lakes free of clouds with colors representing the whole spectrum measured using a standard procedure with a 20 cm diameter white
of lake surface colors in the image. For the sake of quality regres- disk attached to a calibrated depth line. In order to measure chloro-
sion analyses it was important to obtain points with values of phyll a concentration, the water samples from each lake were
bands equally distributed across the whole range of extremal val- passed through Whatman GF/F filters (25 mm diameter) under
ues. The sampled lakes with trophic state classes were determined low suction pressure in order to measure chlorophyll a concen-
J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36 31

Table 1
Modified lake classification scheme (with ranks 1–5) based on trophic state. Ranks 1–5 correspond to Oligotrophic − Eutrophic state. In two bottom rows a number of lakes
in each class is presented for field experiment.

tration [␮g l−1 ] no more than a few hours after collecting. Filters 3.6 ha (mean value = 0.92 ha) depending on the size of the lake and
were kept in a liquid nitrogen until analysis. After thawing, pig- the location of the sampling point.
ments were extracted with the use of 96% ethanol in darkness
at room temperature for 24 h. Then samples were centrifuged for 2.4. Preprocessing of satellite data
15 min at 4000 rpm. The absorbance spectra of the extracts were
determined using a Perkin Elmer Lambda 850 dual-beam spec- Both Landsat 8 scenes (path 190, row 22) were obtained from
trophotometer in a 1-cm cuvette against a 96% ethanol blank. The USGS Glovis and processed in the same way. The DN was first
chlorophyll a concentration was calculated through the use of the converted to TOA reflectance with correction for the sun angle
HELCOM formula (HELCOM, 1988). The chlorophyll a concentra- according to USGS Landsat 8 product instruction (USGS, 2015) using
tion taken into further analysis was calculated as the average value Formula (1). Following this, the atmospheric correction using the
from three aliquots. Suspended particulate matter (SPM; mg/dm3 ) dark object subtraction (DOS) method was carried out assuming 1%
was assessed through vacuum-filtering of 1.8 dm2 water samples surface reflectance from the dark objects (Chavez, 1988) accord-
onto pre-combusted and pre-weighted fibre glass MN GF-5 filters ing to Formulae (2) and (3). The dark objects threshold value is
(mesh size 0.4 ␮m). Large organisms visible to the naked eye were defined as minimum value in particular band. The minimum val-
removed from the filters. Each filter was then air dried at 60 ◦ C ues are most often localized in places with dark land cover (forest)
for 24 h and weighed to determine total suspension dry mass col- covered by cloud shadows or on the surface of small dystrophic or
˛
lected on filters (Zajaczkowski et al., 2010). For concentrations of humic lakes in the forest in cloudless situation. The atmospheric
DOC and CDOM, water from lakes was collected in acid-cleaned correction depends on the band. It changes the blue band (OLI2)
plastic bottles before being transferred to the laboratory within 8 h ten times more than red band (OLI4), for which the correction is
and filtered at low pressure through pre-combusted Whatman GF/F usually insignificant. In result the classification has relative char-
filters. The concentration of dissolved organic carbon (DOC) was acter defined by mean and standard deviation of index value. The
measured with the high temperature catalytic oxidation (HTCO) role of some uncertainty of atmospheric correction is not crucial.
technique using a Vario TOC Cube analyzer (Elementar Analysen- Indeed lack of the correction changes only about 30% of assigning
systeme GmbH).). After filtration, samples were acidified to pH < 2 classes by ± 1 class and only 1% by ± 2 classes. No bias in assigning
with concentrated HCl and stored in a refrigerator until analysis. is observed.
50 ml glass stopper bottles for sample storage were combusted for
5 h at 450 ◦ C. The precision (RSD) of DOC analyses was not worse RTOA = (DN × 0.00002 − 0.1) / cos s (1)
than 3.5%. Absorbance of the filtered water samples was used as a Rscatter = (DNminimum × 0.00002 − 0.1) / cos s (2)
proxy for CDOM concentration. Absorption spectra were recorded
between 250 nm and 700 nm at 1 nm intervals using a Jasco V-630 Rearth = RTOA − (Rscatter − 0.01) (3)
dual-beam spectrophotometer with matching 100 mm quartz cells
where:RTOA − reflectance at the top of atmosphere DN − digital
(10 mm long cell were used when absorbance values exceeded 2.5).
number.  s − solar zenith. Rscatter − reflectance scatter in atmo-
Milli-Q water was used in the reference cell.
sphere. DNminimum − minimum digital number. Rearth − reflectance
Water samples for phytoplankton species and composition were
at earth surface.
immediately preserved in Lugol’s solution. The taxonomic compo-
For bands 1–5, the statistics of corrected reflectance were cal-
sition and number of phytoplankton were analyzed under a NIKON
culated from sets of pixels extracted by polygons surrounding the
inverted microscope. Phytoplankton organisms were identified at
sample points for each lake
the species level or, if not possible, assigned to a genus only. Taxons
were identified using keys and world literature. For each sample,
at least 200 cells were counted with a subset measured to charac- 2.5. Lake classification
terize the size range. Based on cell or colony measurements, algal
biovolume was estimated. After calculating the average volume It was essential to test the relation between spectral signal
of every species, total volume was calculated by multiplying bio- of a lake surface and trophic state of lake. In order to achieve
volume by the number of species. All measurements were saved this, a modified OECD (Vollenweider and Kerekes, 1982) classifi-
as attributes of vector point feature class in the GIS geodatabase cation depending on chl-a and Secchi disk was proposed. Each lake
(ArcMap 10.2). Each point (one sampling point per lake) was sur- included in the field experiment was assigned a rank of 1–5, cor-
rounded by polygons for extraction of statistics of band values from responding with the oligo-to eutrophic state respectively (Fig. 1A).
the raster image. The surfaces of the polygons varied from 0.28 to Ranges of Secchi disk and chlorophyl-a concentration together with
assigning classes are presented in Table 1. Lakes selected for sam-
32 J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36

Table 2 A raster catalog should include all unzipped Landsat 8 bands and
OLI variables candidates for building regression equations with field measurements.
a raster mask of scene with NoData value in areas of clouds and
Variable OLI band combination OLI band combination—description clouds shadows and the value 1 elsewhere. Such a mask must be
VAR1 OLI2 Blue previously prepared for each Landsat 8 scene we want to analyze.
VAR2 OLI3 Green The first step in the model is to calculate reflectance of the first
VAR3 OLI4 Red five bands from the satellite image using the DOS method (Chavez,
VAR4 OLI5 NIR 1988); we must also assume that there is 1% of reflectance from
VAR5 OLI2/OLI3 Blue/Green
the darkest object, as descried earlier. Following this, all reflectance
VAR6 OLI2/OLI4 Blue/Red
VAR7 OLI3/OLI4 Green/Red bands (Ref1,. . .Ref5) are masked using the raster mask. Once this
VAR8 OLI2/OLI5 Blue/NIR has been achieved, for each lake raster zone in the area of the satel-
VAR9 OLI3/OLI5 Green/NIR lite image, the mean values of reflectance of the analyzed bands are
VAR10 (OLI2–OLI4)/(OLI2 + OLI4) (Blue–Red)/(Blue + Red)
calculated using the zonal statistics geoprocessing tool. In addition
VAR11 (OLI2–OLI3)/(OLI2 + OLI3) (Blue–Green)/(Blue + Green)
VAR12 OLI3–OLI2 Green–Blue new raster Ref2/Ref4 is created for which the mean and standard
VAR13 OLI3–OLI4 Green–Red deviation values of reflectance are calculated using also zonal statis-
tics tool. This step creates a table with lake ID (LID) for each lake and
mean values of reflectance for two OLI bands (2,4), Ref2/Ref4 and
pling are shown in Fig. 1A, with trophic state assigned on the basis standard deviation for Ref2/Ref4. The process is controlled by min-
of Secchi disk and Chl-a concentration according to this table. imum pixel number, which is required to calculate statistics. In the
There is a common methodology used when looking at the best next step lakes are divided into subsets on the basis of mean depth.
relation between satellite image information and measured values The initial width of subsets is two meters, with one-meter overlap:
of parameters of water, including trophic state class. Indeed, the 0–2, 1–3, 2–4. . . etc. The one meter overlap is to make classifica-
aforementioned methodology includes testing different variables tion more realistic and smooth. If there are not enough lakes in the
(simple function of satellite image bands) using Ordinary Least subset, everything which is controlled by a parameter – minimum
Squares (OLS) and independent variable transformation. Based on number of lakes in a subset, consecutive subset is merged. The min-
previous research (McCullough et al., 2012; Duan et al., 2008; Sass imum number of lakes in lakes subsets has to be defined because
et al., 2007; Kloiber et al., 2002) several potentially useful Landsat statistics are calculated from each of the subsets. For each subset,
OLI sensor wavelengths were chosen. All tabled (Table 2) variables mean value and standard deviation of trophic state index I is cal-
were used with natural logarithms from all measured water param- culated to show variability of the ecological state of lakes grouped
eters in the OLS model and judged with adjusted R-squared values, into one subset. Following this, thresholds of trophic state index
the basic metric of OLS performance. Additionally, to improve the are used to assign a class to lakes in each subset.These thresholds
method of determining the best combination of bands, we used the are defined as the number of standard deviations of trophic state
significance of each variable defined as the number of times it was index below and above mean index I in each subset. These thresh-
statistically significant with its stability. For all models in the table, olds are common for all subsets. As a result, lakes are divided into
the variable VAR6 (Blue/Red) represents the greatest significance, one of five classes: Bad, Poor, Moderate, Good, High. In the case
with extremely good stability compared to other variables; as such, where a lake occurs in two subsets, it is possible that depth inter-
it was used to calculate the trophic state index. val overlap will occur, potentially resulting in two different classes
being assigned to one lake. If this happens, the better class is cho-
2.6. Model development sen. Defining proper thresholds should be supported with expert
knowledge of the area and may be verified by comparing results
The model (and its implementation as a GIS tool) was created with other classification methods even on the few lakes that have
using two basic assumptions. The first is that there is a variable a diverse ecological state and depth. By comparing the results of
(combination of OLI bands) which correlates well with the trophic different threshold values, models may by calibrated to get results
state of the lakes, while the second is that the reference trophic which best fit the expected classes. It is also possible to test lakes
state depends on the mean depth of the lake. The lake topology for uniformity using coefficient of variation of Ref2/Ref4. For this
trophic state index was introduced by Carlson in 1977 (Carlson, reason standard deviation and mean value is added to results of
1977) on the basis of lake topology (Elster, 1974). The trophic state modeling. The output of the model is a vector layer of lakes with
of freshwater and coastal marine water may be described as olig- an assigned class describing ecological state at the time of a par-
otrophic, mesotrophic, eutrophic and hypertrophic and is related ticular satellite image and mean and standard deviation values of
to amount of total nitrogen, total phosphorus, chlorophyll a and Ref2/Ref4.
Secchi disk transparency (Smith, 1988; Chin, 2006). There were
many projects which proved the relation between remote sensing
reflectance and Chlorophyll a and Secchi disk transparency (Chang 3. Results
et al., 2015) which correlate well with trophic state. The relation-
ship between the depth and size of lakes and its natural trophic 3.1. Field measurements analyses
class is well known in limnology (Lampert and Sommer, 2007). This
relation is explained by several factors. The deep lakes have usually In situ measurements were used to find the best combination of
higher ratio of lake volume to watershed area and in result smaller spectral bands of Landsat 8 satellite images describing the trophic
flux of nutrients per unit lake volume. In deep lakes only a small state of a lake. We used OLS to obtain the linear equation and to
part of epilimnion is in contact with the sediments and in result the predict parameters such as SD or Chl-a from space using a combina-
return of nutrients to the euphotic zones is limited. Large and deep tion of DN, radiance or reflectance. We tested the relation between
lakes have a longer retension time and a large part of nutrients is the natural logarithm of all measured parameters: Secchi depth
lost through sedimentation. (SD), chlorophyll-a (Chl-a), suspended particulate matter (SPM),
There are two inputs into the model, namely the lake geo- colored dissolved organic matter (CDOM) and dissolved organic
database and a raster catalog with Landsat 8 image to be used carbon (DOC) with different variables of spectral band combina-
(Fig. 2). As previously described, the lake geodatabase must con- tions, as shown in Table 2. We used adjusted R Squared (above 0.5 to
tain lake vector layers with mean depths and lake raster zones. accept the model) as well as several diagnostic criteria to evaluate
J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36 33

Fig. 2. The model of classification of lakes to five classes of the trophic state and its implementation in GIS.

Table 3
Results for accepted models. Ln(SPM) − June and Ln(DOC) − June models do not pass tests.

Parameter Scene Adjusted R2 p VAR6 significance [%] VAR6 stability [%]

Ln(SD) June 0.76 <0.05 62 +100


Ln(SD) July 0.81 <0.05 94 +100
Ln(Chl-a) June 0.52 <0.05 31 −100
Ln(Chl-a) July 0.76 <0.05 69 −100
Ln(SPM) June – – 28 −98
Ln(SPM) July 0.77 <0.01 89 −100
Ln(DOC) June – – 42 −96
Ln(DOC) July 0.70 <0.05 67 −100
Ln(CDOM) June 0.60 <0.05 50 −97
Ln(CDOM) July 0.53 <0.05 87 −100

acceptance of the model. Every model was checked for: maximum Table 3. In total there were two sampling campaigns following two
coefficient p-value cutoff (0.05 to accept), maximum VIF value cut- satellite images. However, the attempt to merge data and thus cre-
off (7.5 to accept), minimum acceptable Jarque Bera p-value (0.1 to ate a more general model yielded poorer results, and so models
accept) and minimum acceptable spatial autocorrelation p-value were presented only for separate scenes. For all analyzed variables
(0.1 to accept). Results of the accepted models are presented in VAR6 (Blue/Red) represents the greatest significance, with very
34 J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36

Fig. 3. The relation between measured parameters and VAR 6 (Blue/Red Landsat 8 bands) and mean depth of sampled lakes: a, b,c—SD, Chl-a, tropic class vs VAR 6; d,e,f—SD,
Chl-a, tropic class vs mean depth.

good stability compared to other variables. The significance of this The second part of Fig. 3(d–f) describes the relation of SD, Chl-a
variable was noticed by several authors (Kloiber et al., 2002; Sass and trophic state to mean depth of a lake. Three lakes with mean
et al., 2007; Torbick et al., 2013). depth above 12 m did not fit the general trend and were excluded
The scatterplot analysis (Fig. 3a–c) shows relations described during the creation of the model. Indeed, they have higher Chl-a
by a model between VAR6 and SD, Chl-a and a trophic state rank and lower SD than other lakes with similar mean depth. In lake
defined in Table 1. There is an essential difference in relation for Radunskie Gorne (G), for example, the bloom of diatomophyceae
both scenes. The reason for this may be a result not of the perfect was observed, with biomass volume of 87 000 mm3 /m3 . Both SD
atmospheric corrections or the relatively different composition of and Chl-a are strongly correlated with mean depth (AdjR2 = 0.76
the phytoplankton in June and July. In June diatomphyceae were for SD and AdjR2 = 0.73 for Chl-a) creating a clearly visible trend
present in several lakes while in July they were absent with strong for the relation between trophic state and mean depth (Fig. 3 f)
dominance of chlorophyceae. As a trophic state rank is a function of with AdjR2 = 0.78. This also explains the improvement in results of
SD and Chl-a it correlates well with VAR6 (AdjR2 = 0.68 for June and linear regression achieved by some researchers by including mean
AdjR2 = 0.79 for July with significant p value < 0.01). In conclusion, depth in band variables (McCullough et al., 2012). In our research,
the VAR6 may be treated as a good index of trophic state for single adding mean depth as an additional is not intendent to improve
scene. models for separate scenes. Instead it significantly improves the
general model, making the mean depth the most significant vari-
J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36 35

Fig. 4. The product created by the tool for assigning ecological classes for small part of image area. On the right side the part of image with outlines of lakes and on the left
side the same image with overlaid classified lakes with labels.

Table 4
Verification of the model. Satellite S − September image; Satellite A − April image; best fit − the most fitting choice from Satellite S and A; worst fit − the least fitting choice
from Satellite S and A.

Number of lakes Satellite S/Satellite A Expert/Satellite (best fit) Expert/Satellite (worst fit)

With the same ecological state 9 9 2


With the one class difference 6 9 13
With the two class difference 3 0 3

able. For particular mean depth value, lakes below a regression line (Table 4) shows a good agreement between the satellite and
for trophic class represent a better ecological state than lakes above the expert based method. Half of the lakes have an identical state,
the line. We used this property to relative validate of ecological as indicated by using best fit from both satellite methods and
state of the subset of lakes with similar mean depth. expert judgment. The same number of lakes have identical eco-
logical states, as indicated from both images. There were three
cases of two-class difference noted when comparing both satel-
lite methods and satellite and expert judgment worst scenario. The
3.2. Model implementation
difference was never more than two classes in any comparison and
only one class in the best fit case. The time factor may be critical
The implementation of the tool to assign ecological classes con-
or at least very important when assigning a class to the lake. This,
sists of Python Toolbox with two tools. The first tool has three
from one side, may be treated as a problem, but from the other, may
parameters: Landsat 8 catalog, Landsat 8 MTL text file and the geo-
present the chance to monitor changes in the lakes’ ecological state
database Lake Watch localization. The second tool parameters are:
at the time, by comparing results of models using satellite images
Landsat 8 catalog, localization of Lake Watch geodatabase and the
in time series. This requires further evaluation in the future. We
name of the lake polygon shapefile to be created. In addition, the
also analyzed spatial variation of both bands (OLI2 and OLI4) and
user has to input the value of thresholds between classes Bad and
calculates the coefficient of variation (standard deviation/mean)
Poor, Poor and Moderate, Moderate and Good and Good and High.
for the sample of about 1000 lakes from several images. The mean
The values are the numbers of standard deviations from the mean
value of coefficient of variation was about 0.06 (with standard devi-
values (with minus and plus sign), and may be used for model cal-
ation about 0.05). This shows that reflectance in both bands (as a
ibration. Fig. 4 presents the example of products created using this
trophic state index) has in general low level of spatial variation.
tool. This is only small part of the whole image which was analyzed
However some lakes show higher level of spatial variation. There
to fulfill the requirements of at least 30 lakes in each mean depth
were two reasons for this. The first one is that some shallow lakes
subset.
may be partly covered by reeds during the summer and the second
To verify satellite based methods, we compared results from
is the situation when occasional intensive bloom of phytoplank-
the legal classification of 18 lakes made by experts in 2014 (how-
ton occurs. As a result the coefficient of variation of band 2 or 4
ever some used data were a few years old) with classification of two
may be a good additional index describing the state of the surface
Landsat 8 images (6 September 2014 (S) and 21 April 2015 (A)). The
of particular lake. The values of the coefficient of variation may be
results are shown in Table 4. Parameters used in the model: mini-
calculated for each lake from mean and standard deviation fields
mum number of pixels – 15, minimum number of lakes in subset –
of model output layer. It is possible to define a threshold value to
30, thresholds to differentiate classes: 1.5,0.5,−1,−2 (from High to
inspect some lakes for their uniformity. Also some group analyses
Bad).). The best fit means that the best fitting result from classifica-
may be carried out which will result in dividing lakes to separate
tion of both images has been chosen and the worst fit means that
parts.
the worst result was the choice.
36 J. Andrzej Urbanski et al. / International Journal of Applied Earth Observation and Geoinformation 51 (2016) 28–36

4. Conclusions Carlson, R.E., 1977. A trophic state index for lakes. Limnol. Oceanogr. vol. 22 (2),
361–369.
Chang, N., Imen, S., Vannah, B., 2015. Remote sensing for monitoring surface water
The main limitation in using satellite data to predict SD and Chl- quality status and ecosystem state in relation tot he nutrient cycle: a 40-year
a,e which are considered good indicators of a lake’s trophic state, perspective. Crit. Rev. Environ. Sci. Technol. 45, 101–166.
is that each scene needs its own formula. There are several pos- Chavez, P.S., 1988. An improved dark-object substraction technique for
atmospheric scattering correction of multispectral data. Remote Sens. Environ.
sible reasons for this, but it seems as though the main problem is 24, 459–479.
with the determination of true reflectance (or radiance) due to the Chin, D.A., 2006. Water-Quality Engineering in Natural Systems. John Wiley &
uncertainty with a proper atmospheric correction of the generally Sons, pp. 609.
Duan, H., Zhang, Y., Zhang, B., Song, K., Wang, Z., Liu, D., Li, F., 2008. Estimation of
low signal. The other reasons are due to changing environmental
chlorophyll-a concentration and trophic states for inland lakes in Northeast
conditions: abiotic like lake surface dynamic state or biotic like phy- China from Landsat TM data and field spectral measurements. Int. J. Remote
toplankton composition. It is very important to have representative Sens. 29 (3), 761–786.
Elster, H.J., 1974. History of limnology. Mitteilung Internationale Vereinigung fuer
samples for OLS covering the whole range of trophic states in the
Theoretische unde Amgewandte Limnologie 20, 7–30.
area of research. HELCOM, 1988. Guidelines for the baltic monitoring programme for the third stage
There is strong evidence regarding the statistical relation part D. biological determinands. Baltic Sea Environ. Proc. 27 (D), 153.
between mean depth and expected trophic state of the lake which Janczak, J., 1996–1999. Atlas jezior polski. parts 1,2,3. Poznan, Bogucki
Wydawnictwo Naukowe, 778.
is justified by the functioning of the lake ecosystem. The shallow Kalio, K., 2000. Remote sensing as a tool for monitoring lake water monitoring. In:
lakes do not have constant stratification during the phytoplank- Heinonen, P., Ziglio, G., Van der Beken, A. (Eds.), Hydrological and Limnological
ton growth season, and thus may be effected by resuspension. In Aspects of Lake Monitoring. John Wiley & Sons.
Kloiber, S.M., Brezonik, P.L., Bauer, M.E., 2002. Application of Landsat imagery to
addition to this, the shallow water column may be easily contam- regional-scale assessments of lake clarity. Water Res. 36, 4330–4340.
inated as a result of nutrients delivered by runoff from the land. Lampert, W., Sommer, U., 2007. Limnoecology: The Ecology of Lakes and Streams.
On the other side, deep lakes are free from resuspension and the Oxford University Press, Oxford, pp. 324.
McCullough, I.M., Loftin, C.S., Sader, S.A., 2012. Combining lake and watershed
long lasting stratification and high water column creates a buffer characteristics with Landsat TM data for remote estimation of regional lake
which influences the nutrient download caused by runoff. They are clarity. Remote Sens. Environ. 123, 109–115.
also colder, which slows down biological processes. As all these Olmanson, L.G., Bauer, M.E., Brezonik, P.L., 2008. A 20-year landsat water clarity
census of minnesota’s 10000 lakes. Remote Sens. Environ. 112 (11),
processes are natural, it may be expected that reference lakes will
4086–4097.
have a trophic state depending on their depth. The epilimnion of Peeters, E.T., Franken, R., Jeppesen, E., Moss, B., Becares, E., Hansson, L., Romo, S.,
the lake is an occasionally highly dynamic environment influenced kairesalo, T., Gross, E., Donk, E., Noges, T., Irvine, K., Kornijow, R., Scheffer, M.,
2009. Assessing ecological quality of shallow lakes: does knowledge of
mainly by wind mixing which due to the limited size of the lakes can
transparency suffice? Basic Appl. Ecol. 10, 89–96.
make their surface nearly uniform both physically and biologically. Phillips, G., Pietilainen, O.P., Carvalho, L., Solimini, A., Solheim, L., Cardoso, A., 2008.
However in stable conditions this surface may be less uniform due Chlorophyll-nutrient relationships of different lake types using a large
to blooms or local upwelling events. The mean statistic is the sim- European dataset. Aquat. Ecol. 42, 213–226.
Poikane, S., Portielje, R., Berg, M., Phillips, G., Brucet, S., Carvalho, L., Mischke, U.,
ple generalization of the lake surface state in time scale of weeks. Ott, I., Soszka, H., Wichelen, J., 2014. Defining ecologically relevant water
This assumption may not be however true for very large lakes or quality targets for lakes in Europe. J. Appl. Ecol. 51, 592–602.
coastal waters when their trophic state may differ in various parts. Sass, G.Z., Creed, I.F., Bayley, S.E., Devito, K.J., 2007. Understanding variation in
trophic status of lakes on the Boreal Plain: a 20 year retrospective using
One of the solutions is the subdivision of such waters body. This Landsat TM imagery. Remote Sens. Environ. 109, 127–141.
has been proposed for some lakes in our project. Scheffer, M., 1998. Ecology of Shallow Lakes. Chapman & Hall, London.
The proposed method is an attempt to make use of the increas- Smith, V.H., 1988. Cultural eutrophication of inland, estuarine and coastal waters.
In: Pace, M.L., Groffman, P.M. (Eds.), Success, Limitations and Frontiers in
ing amount of high quality satellite data for the monitoring of vast Ecosystem Science. Springer, New York, pp. 7–49.
numbers of lakes with different trophic states which may change in Song, C., Woodcok, C.E., Karen, C.S., Lenney, M.P., Macomber, S.A., 2001.
the short and long term. As the number of images suitable for such Classification and change detection using landsat TM data: when and how to
correct atmospheric effects? Remote Sens. Environ. 75, 230–244.
analyses will increase with the launching of new satellite systems,
Tebbs, E.J., Remedios, J.j., Harper, D.M., 2013. Remote sensing of chlorophyll-a as a
the statistics of trophic classes for particular lakes may reveal infor- measure of cyanobacterial biomass in Lake Bogoria, a hypertrophic,
mation if its state is stable or changes comparing to other lakes. This saline-alkaline, flamingo lake, using Landsat ETM+. Remote Sens. Environ. 135,
92–106.
may be a valuable tool for the lakes’ environmental managers.
Torbick, N., Hession, S., Hagen, S., Wiangwang, N., Becker, B., Qi, J., 2013. Mapping
inland lake water quality across the Lower Peninsula of Michigan using
Acknowledgments Landsat TM imagery. Int. J. Remote Sens. 34, 7607–7624, 21–22.
USGS, 2015. Landsat 8 (L8), Data Users Handbook, Version 1.0, June 2015, EROS,
Sioux Falls, South Dakota, pp. 98.
The study was financed by the University of Gdansk (530-G300- Vollenweider, R.A., Kerekes, J.J., 1982. Eutrophication of waters Monitoring,
D433-14). We also thank Adam Inglot and Anna Piszewska from assessment and control. In: Final Report. OECD, Paris, pp. 154.
the GIS Center of University of Gdansk for their contribution to the ˛
Zajaczkowski, M., Nygård, H., Hegseth, E.N., Berge, J., 2010. Vertical flux of
particulate matter in an Arctic fjord: the case of lack of the sea-ice cover in
field measurements. Adventfjorden 2006–2007. Polar Biol. 33, 223–239.

References

Brezonik, P., Menken, K.D., Bauer, M., 2005. Landsat-based remote sensing of lake
water quality characteristics, including chlorophyll and colored dissolved
organic matter (CDOM). Lake Reservoir Manage. 21 (4), 373–382.

Intereses relacionados